Artificial grammar learning of shape-based noun classification
Jennifer Culbertson ([email protected])
Linguistics Program, 4400 University Drive
Fairfax, VA 22030 USA
Colin Wilson ([email protected])
Cognitive Science Department, 3400 N. Charles Street
Baltimore, MD 21218 USA
Abstract
Similarly, in the classifier system of Navajo (Mithun, 1986)
nouns are classified according to animacy and shape (among
other properties); class marking in this language is found on
the verb. Signed languages also commonly have noun classification systems based on shape and other functional properties (Supalla, 1986).
Systems of noun classification serve to categorize entities
based on a set of semantic and/or phonological features. Previous work, for the most part focused on gender-based classes,
has suggested that learners acquiring such systems rely primarily on phonological cues, while semantic cues are used
only weakly. We show, using an artificial language learning
task with adults, that semantic information alone is sufficient
to learn a realistic shape-based classification system, challenging the view of phonology bias. Further, our results show that
compared to learners exposed to semantically cohesive categories, learners trained on randomly assigned classes are less
successful at recalling the category of exposure items. This
finding suggests that, contrary to memory-based theories of
learning, categories are not necessarily formed by abstraction
from memorized exemplars, but can instead be constructed
from lower-level properties that category members share.
Keywords: classifiers; noun classes; language acquisition; artificial language learning; semantic features
Acquisition of noun classification systems
Introduction
Systems of noun classification—such as gender, noun class,
and classifier systems—distinguish or categorize objects according to salient semantic and/or phonological features.
Though such systems may differ in their formal realization,
the semantic features on which they are based draw from
a common pool that includes physical features (e.g. shape,
size), function (e.g. food, tool, habitation), as well as animacy
and sociocultural status (Denny, 1976; Dixon, 1986; Lakoff,
1987; Comrie, 1989; Aikhenvald, 2000; Senft, 2000).
For example, in Cantonese, the use of a classifier morpheme is required in constructions involving a numerical or
definite noun phrase, as in example (1) below.1 The choice
of classifier in Cantonese is largely determined by the head
noun; for example the classifier go[3] is used for people,
while the classifier zek[3] is used primarily with animals. Additional classifiers target shape properties like length, dimension, and flexibility.
(1)
a. sam[1] go[3] jan[4]
three CL people
‘three people’
b. sam[1] zek[3] gau[2]
three CL
dogs
‘three dogs’
1 Although English does not use them productively, there are nevertheless a number of nouns which can appear with a classifier, e.g.
“four strands of hair”, “two sheets of paper”, “a school of fish”.
Previous work on the acquisition of systems of noun classification has largely focused on genders and noun classes.
Such studies have documented developmental stages including a period of phonological underspecification, and overgeneralization of frequent or default marking, and have highlighted the apparently weak role of semantic (as opposed
to phonological or distributional) information (KarmiloffSmith, 1981; Perez-Pereira, 1991; Demuth & Ellis, 2008;
Mariscal, 2009; Gagliardi, 2012). The acquisition of classifier systems, although perhaps less well-studied, indicates
a similar developmental trajectory. For example, Tse, Li,
and Leung (2007) report that Cantonese-speaking children
(3;0–5;0) tend to show early use of classifiers in required
contexts but are not adult-like in their choice of classifier
until quite late. In particular, children tend to over-use the
classifier go3—used for people, but also sometimes referred
to as a ‘general’ classifier (C. Li & Thompson, 1989)—and
to over-generalize other more frequent classifiers. Although
P. Li, Huang, and Hsiao (2010) show that Mandarin-speaking
children generalize classifiers to novel nouns on the basis of
shape features, Tsang and Chambers (2011) argue that adult
speakers of Cantonese tend to rely on cues other than the semantic features of the nouns when processing classifiers.
In this paper we investigate the extent to which adult learners can use semantic information alone to acquire category
distinctions instantiated in a miniature classifier system. Previous work on artificial language learning suggests that, although the population of most interest may be children within
any sensitive period for language acquisition, behavioral patterns exhibited by adults can shed light on both general and
language-related learning mechanisms (Wilson, 2006; Culbertson, Smolensky, & Legendre, 2012; Finley & Badecker,
2010). The motivation for using an artificial language learning task rather than natural language learning data in this
case comes from our hypothesis of why it has been found
that phonological cues—even when these are less statistically
reliable than semantic properties—are preferentially used by
2118
learners acquiring noun classification systems (Braine, 1987;
Frigo & McDonald, 1998; Gagliardi, 2012). It seems likely
that children process a great deal of phonological information about dependencies between nouns and nominal modifiers (such as gender-marked determiners or classifiers) before they acquire the meanings of these elements (Polinsky
& Jackson, 1999). In some sense, then, it is unsurprising
that children privilege phonological information at first during language development. Adults may continue to privilege
phonological cues, not because they fail to attend to semantics, but simply because their knowledge of noun classes was
initially based in phonological processing.
Here, crucially, we use adult English-speakers and construct a miniature language from known objects and their linguistic labels. This removes the problem of acquiring the semantics of nouns and, if our hypothesis is correct, should expose an ability to learn cohesive noun categories on the basis
of semantic features alone. While some previous work has
suggested that adults can use semantic information to learn
classification systems in an artificial language, these studies have exclusively focused on gender-based noun classes
(Braine, 1987; Brooks, Braine, Catalano, Brody, & Sudhalter,
1993). Here we target instead shape-based classifiers, which
are likely to be less familiar to English-speaking college students (the population typically targeted).
The system is modeled on Cantonese (sortal) classifiers,
in particular those which pick out shape properties of objects. As mentioned above, the particular shape properties indicated by Cantonese classifiers—related to the length, flexibility, and dimensions of objects—are representative of those
found in classifier systems typologically (Craig, 1986; Dixon,
1986; Comrie, 1989). Table 1 shows the two Cantonese classifiers, along with the semantic features with which they are
associated, on which our system was modeled. The examples
provided represent nouns which take the relevant classifier in
Cantonese, and are also nouns actually used in the task.
random-assignment condition was used to establish an experimental baseline against which performance in the shapebased condition can be compared, and in particular to assess
the role of memory for individual category members in this
task. Exemplar-based models of learning argue that category
formation begins with a set of memorized exemplars, abstract
categories emerging later due to, e.g., computation of featural
similarity among exemplars in a given category (Nosofsky,
1986). This predicts that learners exposed to conditioned
and random classifier categories should perform equally well
when tested on familiar items—in both cases, the set of exemplars presented during exposure should be stored—but should
of course differ on their ability to generalize to novel items.
Participants
Participants were 20 native English-speaking undergraduates
from the Johns Hopkins University. They received a small
amount of course credit or extra credit for their participation.
No subjects reported difficulties hearing or seeing the stimuli.
Materials
The miniature language was comprised of the English numeral words “one” and “two”, two nonce classifier morphemes “ka” and “po”, and 96 English nouns representing
familiar objects. Utterances in the language consisted of a numeral word directly followed by a classifier morpheme, and a
noun, as in example (2) below. Utterances were auditorily—
using mac text-to-speech, speaker “Alex”—and orthographically presented and were accompanied by a visual image.
The image was a single object for numeral “one” or two of
the same objects for numeral “two”.
(2)
b. two-po towel
two-CL towel
Table 1: Shape-based classifiers tested
Classifier
Semantic features
Examples
zi[4]
jeung[4]
rigid, narrow, long
broad, flat, flexible
knife, twig, candle
sheet, card, table
a. one-ka hammer
one-CL hammer
‘two towels’
Experiment 1
In Experiment 1, we tested whether adults could learn and
generalize categories of nouns, distinguished by their use of
the classifiers in Table 1. We compare learning of a system in which classifier use is conditioned on shape-based
semantic properties of nouns to learning a random assignment of nouns to classifier categories. We hypothesized that
if learners perceive and make use of semantic information
in acquiring noun classification systems, they should succeed in learning the semantically-conditioned language. The
Figure 1: Example trial
2119
1.0
0.2
0.4
0.6
0.8
seen
novel
0.0
Participants were seated in front of a computer, and were instructed that the task was about learning a language similar
to English but with two ways of saying the words “one” and
“two”. They then listened to examples of “one-po”, “one-ka”
and “two-po”, “two-ka”. This was followed by 48 familiarization trials, half with objects using the classifier “ka” and
half using “po”. Half of the trials featured a single object
and the other half two objects. On each trial, a visual image
appeared with four choices below it, one for each possible
numeral-classifier combination followed by the object noun
pictured. Participants listened to the auditory stimulus and
were required to click the choice which matched what they
heard. Figure 1 shows an example trial.
After familiarization, participants took a brief break, and
were then instructed that they would see a visual image and
four choices below it, as in the familiarization phase, but they
would hear no audio. Instead they were required to choose
the phrase they thought was most likely to be used in the language. This testing phase was made up of 96 trials, including
all the objects seen during familarization, and 48 novel objects. The seen objects were the same as those seen in the
familiarization phase, but appeared with the other numeral
(e.g. if a participant heard “one-ka hammer” and saw a single
hammer during exposure, they saw two hammers at test). No
feedback was given.
Participants were randomly assigned to one of two conditions. In the shape condition, the use of “ka” and “po”
was conditioned on the semantic properties shown in Table 1
above. The object nouns in each class were a subset of those
which actually use the corresponding classifier in Cantonese.
As such, although they generally exhibited the relevant properties, there was some amount of variation in the extent to
which they did so. For example, the noun “table” takes the
classifier jeung[4] in Cantonese even though it does not perfectly exemplify the semantic features of the class.
In the random condition, the use of “ka” and “po” was unconditioned, and nouns were randomly paired with a particular classifier.
dom effects) reveals a significant effect of condition (β =
1.47, z = 5.32, p < 0.01), with participants in the shape condition choosing the correct classifier on seen items much more
often than those in the random condition (0.86 vs. 0.45).
A significant interaction between condition and number was
also found (β = −0.29, z = −2.63, p < 0.01), indicating the
participants in the random condition tended to be less accurate on items with the number “two” compared to “one”.
We are also interesting in the extent to which participants
in the shape condition could generalize the categorization information they learned during familiarization to novel (unseen) objects at test. As Figure 2 suggests, there was little
difference in participants’ choice of the correct classifier on
seen item, and their choice of the classifier which matched
the relevant semantic features on novel nouns. Analysis using mixed-effects logistic regression revealed no significant
effect of item familiarity (β = 0.27, z = 1.13, p = 0.26). A
significant interaction between item familiarity and number
was found however (β = −0.47, z = −1.98, p < 0.05), indicating the participants tended to be less accurate on seen
items with the number “one” compared to “two”. Note that
for participants in the random condition, there is no expected
correct classifier for novel items, as the noun categories used
in familiarization were random, containing no semantic cues.
Correct response
Design & Procedure
Shape
Random
Condition
Results
In analyzing the results of this experiment we were interested in two main questions: (i) Do learners in the shape
condition—in which classifier choice is determined by semantic features of nouns—succeed in learning and generalizing the correct categories? (ii) Are the categories learned
those which were intended, namely the shape-based categories shown in Table 1? To address the first question, we
compared first the performance on seen items across the two
conditions. Performance on seen items gives an indication
of how well the familiarization set was learned by a given
participant. The light colored bars in Figure 2 shows proportion choice of the correct classifier on average for participants in each condition. Analysis of this data using mixedeffects logistic regression (with participants and items as ran-
Figure 2: Correct choice of classifier for seen and novel nouns
in the shape condition, and seen items in the random condition (NB: there is no correct choice for novel nouns in the
random condition, since nouns were categorized randomly).
If participants in the shape condition in fact consistently
inferred the same set of shape-based categories, we expect to
see that their responses on novel test items are highly correlated. On the other hand, participants in the random condition
were not expected to infer cohesive categories, and thus we
do not expect correlated responses. To assess this, for each
pair of participants in the shape condition, we computed the
proportion of novel test items that they assigned to the same
category. The average agreement proportion for this condi-
2120
1.0
Shape
Random
Condition
Participants
Participants were 24 native English-speaking workers recruited through Amazon Mechanical Turk. They received
$1.00 for their participation in the study.
Figure 3: Correct choice of classifier for seen and novel nouns
in the shape condition, and seen items in the random condition in Experiment 2 (NB: there is no correct choice for novel
nouns in the random condition).
Materials
The materials were the same as those used in Experiment 1,
and participants were again randomly assigned to either the
shape condition or the random condition.
Results
The results of Experiment 2 replicate the major findings of
Experiment 1, as shown in Figure 3. Analysis of this data reveals a significant effect of condition (β = 0.91, z = 3.77, p <
0.01), with participants in the shape condition choosing the
correct classifier on seen items much more often than those
in the random condition (0.82 vs. 0.55). A significant
interaction between condition and number was also found
(β = 0.17, z = 2.08, p < 0.05), indicating that the participants
in the random condition tended to be less accurate on items
with the number “one” compared to “two”. This interaction
is in the opposite direction as what was found in Experiment
1, suggesting that the effect of number may not be reliable.
In terms of generalization to novel items, participants in the
shape condition again show a relatively modest but significant increase in accuracy of classifier choice for seen items in
comparison to novel items (β = 0.38, z = 2.08, p < 0.05). No
other significant effects were observed, again suggesting that
differences in performance driven by number in Experiment
1 may not be reliable.
As in Experiment 1, for each pair of participants in a given
condition we computed the proportion of novel test items that
were assigned to the same category. Average agreement was
above chance for the shape condition (0.65, SE = 0.04), but
note that this represents a lower level of agreement than that
found in Experiment 1 for the same condition. Just as in Experiment 1, average agreement for the random condition was
at the expected chance level (0.50, SE = 0.02).
In the experiments reported above, we exposed adult Englishspeakers to a miniature artificial noun classification system.
0.6
0.0
0.2
In Experiment 2, we sought to replicate our findings in a more
diverse population, namely workers on Amazon Mechanical
Turk (a service pairing workers with tasks over the internet).
This population includes a range of ages and socio-economic
backgrounds that may be more representative of the population at large (Mason & Suri, 2012). In addition, this experiment serves to add to the growing body of linguistic and
cognitive research using Mechanical Turk.
0.4
Correct response
Experiment 2
Discussion
seen
novel
0.8
tion was high (0.74, SE = 0.04). In contrast, a parallel analysis revealed much lower agreement among participants in the
random condition (0.50, SE = 0.02); note that 50% agreement
would be expected from purely random responding.
In order to investigate the role of semantic features of nouns
in the acquisition of classification systems, we used English words, removing an obstacle present in natural language
learning. Child language learners likely go through a stage of
development in which phonological but not semantic information is available for the acquisition of noun classification
and other grammatical features. The results of our experiments indicate that, when exposed to a realistic classification system (based on two Cantonese sortal classifiers) over
known nouns, participants are able to learn the correct categories based on semantic information alone, and can readily
generalize this information to new nouns. Learning did not
extend to participants exposed to randomly generated noun
categories which lacked supporting semantic cues. Our findings were robust in both a population of college students, and
among the more diverse population found on Amazon Mechanical Turk—despite a relatively small sample size.
This finding suggests that semantic features of nouns can
be quickly used by learners as the basis of a classification
system, calling into question the apparently privileged role
of phonology cues argued to hold in previous work on this
topic (Karmiloff-Smith, 1981; Perez-Pereira, 1991; Tsang
& Chambers, 2011; Gagliardi, 2012). While here we have
shown that semantically based noun classification can be
learned in the absence of phonological cues, in future work
we will ask whether phonological information is nevertheless
used preferentially over semantic information when both are
simultaneously accessible.
We believe our results are also relevant to understanding
the initial stages of category formation. In particular, the dramatic difference in performance for seen items—items which
were part of a participant’s exposure set—between the two
conditions calls into question theories of learning in which
2121
categories are formed by abstraction over a set of stored exemplars (Nosofsky, 1986) (see also (Rouder & Ratcliff, 2004)
for relevant discussion and detailed model comparison). Under such a view, the prediction would be that learners should
store the set of exemplars presented during familiarization regardless of whether the particular classifier-noun pairings are
random or semantically conditioned. It would then remain
unexplained why participants in the random condition fail
to use the stored pairings to perform with high accuracy on
seen items at test. Our participants succeeded at remembering (or reconstructing) particular examples only when those
conformed to a more abstract generalization across items.
Acknowledgments
The authors would like to acknowledge Kelly Glazer and Allison Widen for assistance in constructing the experimental
materials and running participants.
References
Aikhenvald, A. (2000). Classifiers: A typology of noun categorization devices. New York, NY: Oxford University
Press.
Braine, M. (1987). What is learned in acquiring word classes:
A step toward an acquisition theory. In B. MacWhinney
(Ed.), Mechanisms of language acquisition (pp. 65–87).
Hillsdale, NJ: Lawrence Erlbaum Associates.
Brooks, P. J., Braine, M. D. S., Catalano, L., Brody, R. E.,
& Sudhalter, V. (1993). Acquisition of gender-like noun
subclasses in an artificial language: The contribution of
phonological markers to learning. Journal of Memory and
Language, 32(1), 76 - 95.
Comrie, B. (1989). Language universals and linguistic typology: Syntax and morphology. Chicago, IL: University
of Chicago press.
Craig, C. (1986). Noun classes and categorization: proceedings of a symposium on categorization and noun classification, Eugene, Oregon, october 1983. Philadelphia, PA:
John Benjamins.
Culbertson, J., Smolensky, P., & Legendre, G. (2012). Learning biases predict a word order universal. Cognition, 122,
306-329.
Demuth, K., & Ellis, D. (2008). Revisiting the acquisition
of Sesotho noun class prefixes. Mahwah, NJ: Lawrence
Erlbaum.
Denny, J. (1976). What are noun classifiers good for? In
Papers from the 12th regional meeting, chicago linguistic
society (pp. 122–132).
Dixon, R. (1986). Noun classes and noun classification in
typological perspective. Noun classes and categorization,
105–112.
Finley, S., & Badecker, W. (2010). Linguistic and nonlinguistic influences on learning biases for vowel harmony.
In S. Ohlsson & R. Catrambone (Eds.), Proceedings of the
32nd annual conference of the Cognitive Science Society
(p. 706-711). Austin, TX: Cognitive Science Society.
Frigo, L., & McDonald, J. (1998). Properties of phonological markers that affect the acquisition of gender-like subclasses. Journal of Memory and Language, 39(2), 218–
245.
Gagliardi, A. (2012). Input and intake in language acquisition. Unpublished doctoral dissertation, PhD dissertation,
University of Maryland.
Karmiloff-Smith, A. (1981). A functional approach to child
language: A study of determiners and reference (Vol. 24).
New York, NY: Cambridge University Press.
Lakoff, G. (1987). Women, fire and dangerous things–what
categories reveal about the mind. Chicago, IL: University
of Chicago Press.
Li, C., & Thompson, S. (1989). Mandarin Chinese: A functional reference grammar. Los Angeles, CA: University of
California press.
Li, P., Huang, B., & Hsiao, Y. (2010). Learning that classifiers
count: Mandarin-speaking children’s acquisition of sortal
and mensural classifiers. Journal of East Asian Linguistics,
19(3), 207–230.
Mariscal, S. (2009). Early acquisition of gender agreement in
the Spanish noun phrase: starting small. Journal of Child
Language, 36(1), 143–171.
Mason, W., & Suri, S. (2012). Conducting behavioral research on Amazon’s Mechanical Turk. Behavior Research
Methods, 44(1), 1–23.
Mithun, M. (1986). The convergence of noun classification
systems. Noun classes and categorization, 379–397.
Nosofsky, R.
(1986).
Attention, similarity, and the
identification–categorization relationship. Journal of Experimental Psychology: General; Journal of Experimental
Psychology: General, 115(1), 39.
Perez-Pereira, M. (1991). The acquisition of gender: What
Spanish children tell us. Journal of Child Language,
18(03), 571–590.
Polinsky, M., & Jackson, D. (1999). Noun classes: Language change and learning. In B. A. Fox, D. Jurafsky, &
L. Michaelis (Eds.), Cognition and function in language
(pp. 29–50). Stanford, CA: CSLI.
Rouder, J., & Ratcliff, R. (2004). Comparing categorization models. Journal of Experimental Psychology: General, 133(1), 63.
Senft, G. (2000). Systems of nominal classification. Cambridge: Cambridge University Press.
Supalla, T. (1986). The classifier system in American sign
language. In C. Craig (Ed.), Noun classes and categorization (pp. 181–214). Philadelphia, PA: John Benjamins.
Tsang, C., & Chambers, C. (2011). Appearances aren’t
everything: Shape classifiers and referential processing in
Cantonese. Journal of Experimental Psychology: Learning, Memory, and Cognition, 37(5), 1065.
Tse, S., Li, H., & Leung, S. (2007). The acquisition of
Cantonese classifiers by preschool children in Hong Kong.
Journal of child language, 34(3), 495.
Wilson, C. (2006). An experimental and computational study
2122
of velar palatalization. Cognitive Science, 30(5), 945-982.
2123